# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import functools from collections.abc import Callable from dataclasses import dataclass from typing import Any, cast import torch from vllm.config import VllmConfig from vllm.distributed import ( get_tensor_model_parallel_world_size, tensor_model_parallel_all_reduce, ) from vllm.forward_context import get_forward_context from vllm.logger import init_logger from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Dynamic from vllm.model_executor.layers.rocm_paged_mxfp4_indexer import ( RocmSparseAttnIndexer, RocmSparseMQAIndexer, ) from vllm.models.deepseek_v41.attention import ( DeepseekV4Attention, DeepseekV4Indexer, DeepseekV4IndexerCache, _replace_layer_index, ) from vllm.models.deepseek_v41.common.ops import ( combine_topk_swa_indices, dequantize_and_gather_k_cache, ) from vllm.models.deepseek_v41.sparse_mla import ( DeepseekV4FlashMLAMetadata, DeepseekV4SparseMLABackend, DeepseekV4SparseMLAMetadataBuilder, DeepseekV41SparseSWAMetadataBuilder, ) from vllm.platforms import current_platform from vllm.platforms.rocm import _ON_GFX950 from vllm.triton_utils import tl, triton from vllm.utils.multi_stream_utils import execute_in_parallel from vllm.v1.attention.backend import ( AttentionCGSupport, CommonAttentionMetadata, MultipleOf, ) from vllm.v1.attention.backends.mla.sparse_swa import ( DeepseekSparseSWABackend, DeepseekSparseSWAMetadata, ) from vllm.v1.attention.ops.rocm_aiter_mla_sparse import ( build_prefill_topk_ragged_indices, build_ragged_indices_from_dense, rocm_inv_rope_einsum, rocm_inverse_rope_mxfp8_rows, rocm_inverse_rope_rows_, rocm_mxfp8_wo_a_bmm, rocm_sparse_attn_decode, rocm_sparse_attn_prefill, ) from vllm.v1.attention.ops.rocm_paged_mxfp4_indexer import ( rocm_paged_mxfp4_cache_layout, ) from vllm.v1.kv_cache_interface import KVCacheSpec from vllm.v1.worker.workspace import current_workspace_manager logger = init_logger(__name__) def _trust_dsv4_extra_cache_nan_free( kv_cache_dtype: str, has_kv_transfer: bool, has_extra_cache: bool, ) -> bool: return ( _ON_GFX950 and kv_cache_dtype == "fp8_ds_mla" and not has_kv_transfer and has_extra_cache ) def _aiter_sparse_mla_enabled(has_kv_transfer: bool) -> bool: from vllm._aiter_ops import rocm_aiter_ops if not rocm_aiter_ops.is_triton_sparse_mla_enabled(): return False # The aiter kernel does not scrub NaNs, so like the trust check above it # needs a cache only the canonical writer has touched. if has_kv_transfer: logger.warning_once( "VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA is ignored with a KV " "connector; DeepSeek V4 keeps the Triton sparse attention." ) return False return True def _build_indptr_from_lengths(lengths: torch.Tensor) -> torch.Tensor: lengths = lengths.to(dtype=torch.int32).contiguous() indptr = torch.zeros(lengths.shape[0] + 1, dtype=torch.int32, device=lengths.device) torch.cumsum(lengths, dim=0, out=indptr[1:]) return indptr def apply_pre_quantized_block_scaled_mm( linear: torch.nn.Module, x_fp8: torch.Tensor, x_scale: torch.Tensor, ) -> torch.Tensor: """Block-scaled fp8 GEMM on pre-quantized activations. The fused q/kv norm kernel writes fp8 qr + per-1x128 scales; this drives the linear's block-scaled GEMM directly with them, bypassing apply_weights which would re-quantize the fp8 input. Only valid for the wq_b-style column/replicated linears: their output is the local TP shard, so no all-reduce is needed. """ from vllm.model_executor.kernels.linear.scaled_mm.BlockScaledMMLinearKernel import ( FP8BlockParams, ) params = FP8BlockParams.from_layer(linear) weight_scale = ( params.weight_scale if params.weight_scale_inv is None else params.weight_scale_inv ) kernel = linear.quant_method.fp8_linear out = kernel.apply_block_scaled_mm( A=x_fp8, B=params.weight, As=x_scale, Bs=weight_scale ) return out.to(dtype=kernel.config.out_dtype) @triton.jit def _compute_topk_lens_kernel( topk_lens_ptr, topk_indices_ptr, topk_indices_stride, topk, is_valid_token_ptr, TRITON_BLOCK_SIZE: tl.constexpr, ): token_idx = tl.program_id(0) is_valid_token = tl.load(is_valid_token_ptr + token_idx) count = tl.zeros((), dtype=tl.int32) for i in range(0, topk, TRITON_BLOCK_SIZE): offset = i + tl.arange(0, TRITON_BLOCK_SIZE) mask = offset < topk local_idx = tl.load( topk_indices_ptr + token_idx * topk_indices_stride + offset, mask=mask, other=-1, ) count += tl.sum((local_idx >= 0).to(tl.int32), axis=0) tl.store(topk_lens_ptr + token_idx, tl.where(is_valid_token, count, 0)) @triton.jit def _pack_global_topk_ragged_kernel( global_topk_ragged_ptr, topk_indptr_ptr, topk_indices_ptr, topk_indices_stride, token_to_req_indices_ptr, block_table_ptr, block_table_stride, block_size, topk, BLOCK_SIZE: tl.constexpr, ): token_idx = tl.program_id(0) block_idx = tl.program_id(1) offset = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) out_start = tl.load(topk_indptr_ptr + token_idx) out_end = tl.load(topk_indptr_ptr + token_idx + 1) out_len = out_end - out_start if block_idx * BLOCK_SIZE >= out_len: return req_idx = tl.load(token_to_req_indices_ptr + token_idx) mask = (offset < out_len) & (offset < topk) local_idx = tl.load( topk_indices_ptr + token_idx * topk_indices_stride + offset, mask=mask, other=-1, ) valid = mask & (local_idx >= 0) block_indices = local_idx // block_size block_numbers = tl.load( block_table_ptr + req_idx * block_table_stride + block_indices, mask=valid, other=0, ) block_offsets = local_idx % block_size slot_ids = tl.where(valid, block_numbers * block_size + block_offsets, -1) tl.store(global_topk_ragged_ptr + out_start + offset, slot_ids, mask=mask) def compute_global_topk_ragged_indices_and_indptr( topk_indices: torch.Tensor, token_to_req_indices: torch.Tensor, block_table: torch.Tensor, block_size: int, is_valid_token: torch.Tensor, ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: topk_indices = topk_indices.reshape(topk_indices.shape[0], -1).contiguous() num_tokens = topk_indices.shape[0] topk = topk_indices.shape[1] topk_lens = torch.empty(num_tokens, dtype=torch.int32, device=topk_indices.device) _compute_topk_lens_kernel[(num_tokens,)]( topk_lens, topk_indices, topk_indices.stride(0), topk, is_valid_token, TRITON_BLOCK_SIZE=1024, ) topk_indptr = _build_indptr_from_lengths(topk_lens) global_topk_ragged = torch.empty( num_tokens * topk, dtype=torch.int32, device=topk_indices.device, ) if global_topk_ragged.numel() > 0: block = 128 _pack_global_topk_ragged_kernel[(num_tokens, triton.cdiv(topk, block))]( global_topk_ragged, topk_indptr, topk_indices, topk_indices.stride(0), token_to_req_indices, block_table, block_table.stride(0), block_size, topk, BLOCK_SIZE=block, ) return global_topk_ragged, topk_indptr, topk_lens def _copy_ragged_to_graph_buffers( ragged_indices: torch.Tensor, ragged_indptr: torch.Tensor, ragged_indices_buffer: torch.Tensor, ragged_indptr_buffer: torch.Tensor, num_rows: int, max_entries_per_row: int, ) -> tuple[torch.Tensor, torch.Tensor]: """Copy dynamic ragged metadata into persistent CUDA graph buffers. FULL decode graphs capture kernel argument addresses. Keep the returned tensors backed by stable storage, while indptr continues to bound reads. """ indptr_out = ragged_indptr_buffer[: num_rows + 1] indptr_out.copy_(ragged_indptr, non_blocking=True) max_entries = max(num_rows * max_entries_per_row, 1) ragged_out = ragged_indices_buffer[:max_entries] source_entries = ragged_indices.numel() if source_entries > 0: ragged_out[:source_entries].copy_(ragged_indices, non_blocking=True) if _ON_GFX950: # Preserve the graph-stable base pointer while exposing source capacity # to the sync-free split selector; indptr still carries the true NNZ. ragged_out = ragged_out[: max(source_entries, 1)] return ragged_out, indptr_out # (ragged_indices, ragged_indptr, lens) as consumed by rocm_sparse_attn_decode. _TopkRagged = tuple[torch.Tensor, torch.Tensor, torch.Tensor] @dataclass class DeepseekV41ROCMAiterSparseSWAMetadata(DeepseekSparseSWAMetadata): decode_swa_ragged_indices: torch.Tensor | None = None decode_swa_ragged_indptr: torch.Tensor | None = None # Built only for the aiter prefill, which reads the SWA cache in place. prefill_swa_ragged_indices: torch.Tensor | None = None prefill_swa_ragged_indptr: torch.Tensor | None = None class DeepseekV41ROCMAiterMLASparseMetadataBuilder(DeepseekV4SparseMLAMetadataBuilder): @classmethod def get_cudagraph_support( cls, vllm_config: VllmConfig, kv_cache_spec: KVCacheSpec, ) -> AttentionCGSupport: spec_config = vllm_config.speculative_config if spec_config is not None and spec_config.enable_adaptive_verification: # Per-token request ownership and compressed slot mappings are built # from device query boundaries into persistent buffers. return AttentionCGSupport.ALWAYS return super().get_cudagraph_support(vllm_config, kv_cache_spec) class DeepseekV41ROCMAiterSparseSWAMetadataBuilder(DeepseekV41SparseSWAMetadataBuilder): @classmethod def get_cudagraph_support( cls, vllm_config: VllmConfig, kv_cache_spec: KVCacheSpec, ) -> AttentionCGSupport: spec_config = vllm_config.speculative_config if spec_config is not None and spec_config.enable_adaptive_verification: # SWA indices, lengths, and request ownership are built from device # query boundaries and copied into persistent ragged buffers. return AttentionCGSupport.ALWAYS return super().get_cudagraph_support(vllm_config, kv_cache_spec) # Keep fused multi-step decode disabled until update_draft_decode_metadata() # also refreshes the ROCm-specific ragged SWA indices and indptrs. supports_draft_decode_metadata_update = False def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) max_tokens = self.vllm_config.scheduler_config.max_num_batched_tokens # The non-causal (DSpark draft) path widens each token's SWA index list # to ``noncausal_index_width`` (>= window_size), so size the persistent # ragged buffer to the wider bound to cover both causal and non-causal. swa_index_width = max(self.window_size, self.noncausal_index_width) self.decode_swa_ragged_indices_buffer = torch.empty( max_tokens * swa_index_width, dtype=torch.int32, device=self.device, ) self.decode_swa_ragged_indptr_buffer = torch.empty( max_tokens + 1, dtype=torch.int32, device=self.device, ) self._use_aiter_sparse_mla = _aiter_sparse_mla_enabled( self.vllm_config.kv_transfer_config is not None ) def build( self, common_prefix_len: int, common_attn_metadata: CommonAttentionMetadata, fast_build: bool = False, replay_start: torch.Tensor | None = None, ) -> DeepseekV41ROCMAiterSparseSWAMetadata: base = super().build( common_prefix_len=common_prefix_len, common_attn_metadata=common_attn_metadata, fast_build=fast_build, replay_start=replay_start, ) prefill_ragged_indices = None prefill_ragged_indptr = None if ( self._use_aiter_sparse_mla and base.num_prefill_tokens > 0 and base.prefill_swa_indices is not None and base.prefill_swa_lens is not None ): # Paged-coordinate rows, so every layer's prefill shares one build. prefill_ragged_indices, prefill_ragged_indptr = ( build_ragged_indices_from_dense( base.prefill_swa_indices.reshape(base.num_prefill_tokens, -1), base.prefill_swa_lens, ) ) ragged_indices = None ragged_indptr = None if ( base.num_decode_tokens > 0 and base.decode_swa_indices is not None and base.decode_swa_lens is not None ): ragged_indices, ragged_indptr = build_ragged_indices_from_dense( base.decode_swa_indices.reshape( base.num_decode_tokens, base.decode_swa_width ), base.decode_swa_lens, ) ragged_indices, ragged_indptr = _copy_ragged_to_graph_buffers( ragged_indices, ragged_indptr, self.decode_swa_ragged_indices_buffer, self.decode_swa_ragged_indptr_buffer, base.num_decode_tokens, base.decode_swa_width, ) return DeepseekV41ROCMAiterSparseSWAMetadata( **vars(base), decode_swa_ragged_indices=ragged_indices, decode_swa_ragged_indptr=ragged_indptr, prefill_swa_ragged_indices=prefill_ragged_indices, prefill_swa_ragged_indptr=prefill_ragged_indptr, ) @functools.cache def _aiter_indexer_cache_ops() -> tuple[Callable[..., None], Callable[..., tuple]]: """The aiter indexer key writer and query quantizer.""" from aiter.ops.triton.fusions.k_norm_rope_mxfp4_cache import ( k_norm_rope_mxfp4_cache, ) from aiter.ops.triton.rope.q_rope_mxfp4_quant import q_rope_mxfp4_quant return k_norm_rope_mxfp4_cache, q_rope_mxfp4_quant def rocm_mxfp4_indexer_k_store( k_pre: torch.Tensor, positions: torch.Tensor, cos_sin_cache: torch.Tensor, rms_norm_weight: torch.Tensor, rms_norm_eps: float, k_cache: torch.Tensor, kv_slot_mapping: torch.Tensor, compress_ratio: int, use_fp4_cache: bool, *, num_heads: int, ) -> None: """`indexer_k_norm_rope_store` for the ROCm MXFP4 cache: aiter's cache op writes the key in the order its MQA-logits kernel reads with ``num_heads`` query heads.""" assert use_fp4_cache, "the ROCm indexer cache op writes MXFP4 only" layout = rocm_paged_mxfp4_cache_layout(num_heads, k_pre.shape[1], k_cache.shape[1]) k_norm_rope_mxfp4_cache, _ = _aiter_indexer_cache_ops() k_norm_rope_mxfp4_cache( k_pre, positions, cos_sin_cache, rms_norm_weight, rms_norm_eps, k_cache, kv_slot_mapping, compress_ratio, shuffle=layout, ) def rocm_mxfp4_indexer_q_quant( positions: torch.Tensor, index_q: torch.Tensor, index_q_cos_sin_cache: torch.Tensor, index_weights: torch.Tensor, index_weights_softmax_scale: float, index_weights_head_scale: float, use_fp4: bool = True, weights_out_dtype: torch.dtype = torch.float32, ) -> tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor]: """`fused_indexer_q_rope_quant` for the ROCm MXFP4 indexer, on aiter's op: ((packed [T, H, D // 2], e8m0 as one int32 per head [T, H]), fp32 weights).""" assert use_fp4 and weights_out_dtype == torch.float32, ( "the ROCm indexer quantizes Q to MXFP4 and scores with fp32 weights" ) _, q_rope_mxfp4_quant = _aiter_indexer_cache_ops() q_packed, q_scale, weights_out = q_rope_mxfp4_quant( index_q, positions, index_q_cos_sin_cache, index_weights, index_weights_softmax_scale * index_weights_head_scale, ) return (q_packed, q_scale.view(torch.int32).squeeze(-1)), weights_out class DeepseekV41RocmMxfp4Indexer(DeepseekV4Indexer): """The indexer on aiter's paged MXFP4 cache: its K store and Q quant write in the order aiter's MQA-logits kernel reads, and its layers score with it. ``_produce_k`` and ``forward_q`` mirror ``DeepseekV4Indexer``'s except for those two calls.""" mqa_cls = RocmSparseMQAIndexer attn_cls = RocmSparseAttnIndexer def _produce_k( self, latent: torch.Tensor | None, positions: torch.Tensor, rotary_emb: torch.nn.Module, ) -> None: attn_metadata = get_forward_context().attn_metadata if not isinstance(attn_metadata, dict) or latent is None: return assert self.owns_k indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix]) k_pre, _ = self.wk(latent) rocm_mxfp4_indexer_k_store( k_pre, positions, rotary_emb.cos_sin_cache, self.k_norm.weight, self.k_norm.variance_epsilon, self.k_cache.kv_cache, indexer_metadata.slot_mapping, self.compress_ratio, self.use_fp4_kv, num_heads=self.n_head, ) def forward_q( self, qr: torch.Tensor | QuantizedActivation, qr_scale: torch.Tensor | None, indexer_weights: torch.Tensor, positions: torch.Tensor, rotary_emb: torch.nn.Module, ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]: q = self._wq_b_proj(qr, qr_scale).view(-1, self.n_head, self.head_dim) (q, q_scale), weights = rocm_mxfp4_indexer_q_quant( positions, q, rotary_emb.cos_sin_cache, indexer_weights, self.softmax_scale, self.n_head**-0.5, use_fp4=self.use_fp4_kv, weights_out_dtype=self.indexer_weights_dtype, ) return q, q_scale, weights class DeepseekV41ROCMAiterMLASparseBackend(DeepseekV4SparseMLABackend): @staticmethod def get_name() -> str: return "ROCM_FLASHMLA_SPARSE_DSV41" @staticmethod def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: return [64, 128] @staticmethod def get_builder_cls() -> type[DeepseekV41ROCMAiterMLASparseMetadataBuilder]: return DeepseekV41ROCMAiterMLASparseMetadataBuilder class DeepseekV41ROCMAiterSparseSWABackend(DeepseekSparseSWABackend): @staticmethod def get_builder_cls() -> type["DeepseekV41ROCMAiterSparseSWAMetadataBuilder"]: return DeepseekV41ROCMAiterSparseSWAMetadataBuilder class DeepseekV41ROCMAiterMLAAttention(DeepseekV4Attention): """ROCm sparse MLA attention layer for DeepSeek V4.1.""" backend_cls = DeepseekV41ROCMAiterMLASparseBackend swa_backend_cls = DeepseekV41ROCMAiterSparseSWABackend _use_aiter_sparse_mla = False def _indexer_cls( self, k_cache: DeepseekV4IndexerCache | None ) -> type[DeepseekV4Indexer]: if k_cache is not None and k_cache.rocm_mxfp4: return DeepseekV41RocmMxfp4Indexer return super()._indexer_cls(k_cache) def __init__(self, *args, **kwargs): vllm_config = args[0] if args else kwargs["vllm_config"] super().__init__(*args, **kwargs) # CUDA executes WO_A with a quantized grouped-BMM kernel. ROCm's # correctness path below dequantizes WO_A once and uses torch.einsum, # so retain the ordinary MXFP8 linear kernel for post-load processing # instead of asking for the CUDA-only BMM kernel. self.wo_a.is_bmm = False self._has_kv_transfer = vllm_config.kv_transfer_config is not None # Block scale for the preshuffled weight; None = not preshuffled. self._wqa_wkv_scale: torch.Tensor | None = None self._wo_b_scale: torch.Tensor | None = None self._fused_compressor_weight: torch.Tensor | None self.register_buffer("_fused_compressor_weight", None, persistent=False) self._fused_compressor_split_sizes: tuple[int, int] | None = None # Decode ragged topk metadata is a pure function of the indices its # index source published, so every consumer of a source rebuilds the # same thing. The source owns one cache per compress ratio, refreshed # when it runs; consumers below it read through. self._topk_ragged_cache: dict[int, _TopkRagged] = {} self._prefill_topk_ragged_cache: dict[ str, tuple[torch.Tensor, torch.Tensor] ] = {} self._index_source_prefix: str | None = None if self.compress_ratio > 0: assert self.index_source_layer_id is not None self._index_source_prefix = _replace_layer_index( self.prefix, self.index_source_layer_id ) if self._index_source_prefix not in self._static_forward_context: raise NotImplementedError( f"Index source {self._index_source_prefix} not found on " "this rank; PP splits inside a v4.1 index-sharing group " "are not supported." ) self._use_aiter_sparse_mla = _aiter_sparse_mla_enabled(self._has_kv_transfer) if self.compressor is None and self.indexer is None: # Dense layers have nothing to overlap; keep the base serial path. self.aux_stream_list = None @classmethod def get_padded_num_q_heads(cls, num_heads: int) -> int: return num_heads def prepare_attn_preshuffle(self) -> None: from vllm._aiter_ops import rocm_aiter_ops if not rocm_aiter_ops.is_enabled(): return from vllm.model_executor.layers.quantization.utils.fp8_utils import ( _upcast_e8m0_to_fp32, ) from vllm.model_executor.utils import replace_parameter def _prep(linear) -> torch.Tensor | None: w = getattr(linear, "weight", None) if w is None or w.dim() != 2: return None # K % 128 (group-128 quant) and N % 16 (shuffle_weight) must hold. if w.shape[-1] % 128 != 0 or w.shape[0] % 16 != 0: return None ws = getattr(linear, "weight_scale_inv", None) # per-block scale if ws is None: return None if ws.dtype == torch.float8_e8m0fnu: ws = _upcast_e8m0_to_fp32(ws).contiguous() # Shuffle the weight in place (single weight, no unshuffled copy). replace_parameter( linear, "weight", rocm_aiter_ops.shuffle_weight(w.data, layout=(16, 16)), ) return ws self._wqa_wkv_scale = _prep(self.fused_wqa_wkv) self._wo_b_scale = _prep(self.wo_b) def prepare_compressor_gemm_fusion(self) -> bool: # V4.1 derives index keys from the source compressor's emitted latent # and has no nested ``indexer.compressor``. Keep the projections # separate and use the shared linear/PyTorch correctness path. return False def _bpre_attn_gemm( self, weight: torch.Tensor, scale: torch.Tensor, x: torch.Tensor, reduce_tp: bool, ) -> torch.Tensor: from vllm._aiter_ops import rocm_aiter_ops x_fp8, x_scale = rocm_aiter_ops.group_fp8_quant(x, transpose_scale=True) out = rocm_aiter_ops.gemm_a8w8_blockscale_bpreshuffle( x_fp8, weight, x_scale, scale, output_dtype=x.dtype ) if reduce_tp and get_tensor_model_parallel_world_size() > 1: out = tensor_model_parallel_all_reduce(out) return out def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor: if self._wqa_wkv_scale is not None and hidden_states.dim() == 2: return self._bpre_attn_gemm( self.fused_wqa_wkv.weight, self._wqa_wkv_scale, hidden_states, False ) return super()._fused_wqa_wkv_gemm(hidden_states) def _run_parallel_input_projections( self, hidden_states: torch.Tensor ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]: if self.indexer is not None and self.compressor is None: # Serial: qr must sit on the default stream before the q-side # fork below can consume it (ROCm). aux_streams = self.aux_stream_list self.aux_stream_list = None try: return super()._run_parallel_input_projections(hidden_states) finally: self.aux_stream_list = aux_streams return super()._run_parallel_input_projections(hidden_states) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, llama_4_scaling: torch.Tensor | None = None, ) -> torch.Tensor: if ( self.indexer is None or self._prepare_and_attn_fn == self._prepare_and_attn_eager or not current_platform.enable_multi_stream_overlap( self.aux_stream_list, get_forward_context().attn_metadata ) ): # Sequential fallback: no forks outside the capture region, where # HIP event sync is unreliable; MRV1 keeps the input prep in its # wide eager region (#51430). aux_streams = self.aux_stream_list self.aux_stream_list = None try: return super().forward(positions, hidden_states, llama_4_scaling) finally: self.aux_stream_list = aux_streams attn_out = self._alloc_attn_out(hidden_states.shape[0], hidden_states) if self.compressor is not None: q, kv, index_q, index_q_scale, index_weights = self._forward_csa2_full( hidden_states, positions ) else: q, kv, index_q, index_q_scale, index_weights = self._forward_csa2_reindex( hidden_states, positions ) self._sparse_indexer_and_attn( hidden_states, index_q, index_q_scale, index_weights, q, kv, positions, attn_out, ) return self._o_proj(attn_out, positions) def _forward_csa2_full( self, hidden_states: torch.Tensor, positions: torch.Tensor ) -> tuple[ torch.Tensor, torch.Tensor, torch.Tensor | None, torch.Tensor | None, torch.Tensor | None, ]: """CSA layers: one fork replaces the base pipeline's three. Each fork/join is a HIP event pair per side stream on ROCm, so the base three-fork pipeline is expensive there; the compressor chain and the indexer weights projection run fully in parallel with the default chain. The indexer K write and q-side stay serial after the join: ``wk`` consumes the aux-0 latent and the q-side's wq_b consumes the default-stream qr. """ attn_metadata = get_forward_context().attn_metadata compressor = self.compressor indexer = self.indexer assert compressor is not None and indexer is not None aux_streams = self.aux_stream_list assert aux_streams is not None and len(aux_streams) >= 2 def default_chain(): qr_kv = self._fused_wqa_wkv_gemm(hidden_states) qr, qr_scale, kv = self._split_qkv_and_norm(qr_kv) q = self._wq_b_proj(qr, qr_scale).view( -1, self.n_local_heads, self.head_dim ) return ( self._fused_qnorm_rope_kv_insert(q, kv, positions, attn_metadata), qr, qr_scale, kv, ) def compressor_chain(): kv_score = torch.mm( hidden_states, compressor.fused_wkv_wgate.weight.T, out_dtype=torch.float32, ) latent = compressor(kv_score, positions) compressor.insert_cache(latent, positions, self.rotary_emb) return latent def indexer_weights_chain(): # ReplicatedLinear returns (output, bias); bias is None. weights, _ = indexer.weights_proj(hidden_states) return weights (q, qr, qr_scale, kv), (latent, indexer_weights) = execute_in_parallel( default_chain, [compressor_chain, indexer_weights_chain], self.ln_events[0], self.ln_events[1:3], aux_streams[:2], enable=True, default_first=True, ) indexer._produce_k(latent, positions, self.indexer_rotary_emb) index_q, index_q_scale, index_weights_out = indexer.forward_q( qr, qr_scale, indexer_weights, positions, self.indexer_rotary_emb ) return q, kv, index_q, index_q_scale, index_weights_out def _forward_csa2_reindex( self, hidden_states: torch.Tensor, positions: torch.Tensor ) -> tuple[ torch.Tensor, torch.Tensor, torch.Tensor | None, torch.Tensor | None, torch.Tensor | None, ]: """Indexer-only layers: the q-side overlaps the SWA q path. The input projections run serial first: the q-side consumes qr, so it can only overlap the SWA q projection and KV insert. """ indexer = self.indexer assert indexer is not None attn_metadata = get_forward_context().attn_metadata # Serial on these layers via the _run_parallel_input_projections # override: qr must sit on the default stream before the fork. qr_kv, _, indexer_weights = self._run_parallel_input_projections(hidden_states) assert indexer_weights is not None qr, qr_scale, kv = self._split_qkv_and_norm(qr_kv) def swa_q_chain(): q = self._wq_b_proj(qr, qr_scale).view( -1, self.n_local_heads, self.head_dim ) return self._fused_qnorm_rope_kv_insert(q, kv, positions, attn_metadata) def indexer_q_chain(): return indexer.forward_q( qr, qr_scale, indexer_weights, positions, self.indexer_rotary_emb ) aux_streams = self.aux_stream_list assert aux_streams is not None q, aux_results = execute_in_parallel( swa_q_chain, [indexer_q_chain], self.ln_events[0], self.ln_events[1:2], aux_streams[:1], enable=True, default_first=True, ) index_q, index_q_scale, index_weights_out = aux_results[0] return q, kv, index_q, index_q_scale, index_weights_out @functools.cached_property def _wq_b_uses_aiter_block_scaled(self) -> bool: """True when both wq_b GEMMs run the aiter block-scaled fp8 kernel. Cached: the linear kernels and the aiter env gates are fixed once the model is built, so this is evaluated at the first forward only. The fused norm+quant path is only valid if the quant and GEMM it replaces are exactly the aiter ones; otherwise fall back to the shared path. """ from vllm._aiter_ops import rocm_aiter_ops from vllm.model_executor.kernels.linear.scaled_mm import ( Fp8BlockScaledMMLinearKernel, ) if not rocm_aiter_ops.is_linear_fp8_enabled(): return False linears = [self.wq_b] if self.indexer is not None: linears.append(self.indexer.wq_b) for linear in linears: kernel = getattr(getattr(linear, "quant_method", None), "fp8_linear", None) if not isinstance(kernel, Fp8BlockScaledMMLinearKernel): return False return True def _split_qkv_and_norm( self, qr_kv: torch.Tensor ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor]: """Fuse q/kv RMSNorm + per-1x128 fp8 q quant into one aiter kernel. The shared path norms q and kv in one triton kernel and the wq_b linears then re-read the bf16 qr to quantize it. The aiter kernel computes both RMSNorms (fp32 accumulate) and the fp8 group quant in a single pass, writing fp8 qr + group scales directly; both wq_b GEMMs (attention and indexer) then consume that pair and skip their own input quant. kv stays bf16: the fused insert kernel RoPE/quantizes it itself. Falls back to the shared path when the aiter linear path is not active. """ qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1) if not ( qr.dim() == 2 and qr.shape[0] > 0 and self.q_lora_rank % 128 == 0 and self._wq_b_uses_aiter_block_scaled ): return super()._split_qkv_and_norm(qr_kv) from vllm._aiter_ops import rocm_aiter_ops return rocm_aiter_ops.fused_qk_rmsnorm_group_quant( q=qr, q_weight=self.q_norm.weight.data, q_epsilon=self.eps, kv=kv, kv_weight=self.kv_norm.weight.data, kv_epsilon=self.eps, group_size=128, transpose_scale=False, ) def _alloc_attn_out( self, num_tokens: int, hidden_states: torch.Tensor ) -> torch.Tensor | QuantizedActivation: if not _ON_GFX950: return super()._alloc_attn_out(num_tokens, hidden_states) # wo_a's MXFP8 input: the decode reduce writes it directly, prefill # rows are rotated and quantized after their bf16 attention. width = self.n_local_heads * self.head_dim data = torch.empty( (num_tokens, width), dtype=torch.float8_e4m3fn, device=hidden_states.device ) scale = torch.empty( (num_tokens, width // 32), dtype=torch.uint8, device=hidden_states.device ) return QuantizedActivation( data=data, scale=scale, orig_dtype=hidden_states.dtype, orig_shape=data.shape, quant_key=kMxfp8Dynamic, ) def _o_proj( self, attn_out: torch.Tensor | QuantizedActivation, positions: torch.Tensor ) -> torch.Tensor: if isinstance(attn_out, QuantizedActivation): z = rocm_mxfp8_wo_a_bmm( attn_out.data, attn_out.scale, self.wo_a, self.n_local_groups, self.o_lora_rank, ) return self._wo_b_after_wo_a(z) o = attn_out[:, : self.n_local_heads, :] # ROCm BF16 reference wo_a path (inverse RoPE + einsum) + wo_b. z = rocm_inv_rope_einsum( self.rotary_emb, o, positions, self.rope_head_dim, self.n_local_groups, self.o_lora_rank, self.wo_a, inverse_rope=False, ) return self._wo_b_after_wo_a(z.flatten(1)) def _wo_b_after_wo_a(self, zf: torch.Tensor) -> torch.Tensor: if self._wo_b_scale is not None and zf.dim() == 2: return self._bpre_attn_gemm( self.wo_b.weight, self._wo_b_scale, zf, self.wo_b.reduce_results ) return self.wo_b(zf) def forward_mqa( self, q: torch.Tensor, kv: torch.Tensor, positions: torch.Tensor, output: torch.Tensor | QuantizedActivation, ) -> None: mxfp8_out = output if isinstance(output, QuantizedActivation) else None if mxfp8_out is None: assert isinstance(output, torch.Tensor) assert output.shape == q.shape, ( f"output buffer shape {output.shape} must match q shape {q.shape}" ) assert output.dtype == q.dtype, ( f"output buffer dtype {output.dtype} must match q dtype {q.dtype}" ) forward_context = get_forward_context() attn_metadata = forward_context.attn_metadata if attn_metadata is None: # Warmup dummy run: no real metadata. Reserve the workspace # _forward_prefill would; the dequantize / topk / sparse_fwd kernels # are skipped this step. The aiter prefill reads the caches in place, # so it asks only for the bf16 rows of the MXFP8 output. swa_only = self.compress_ratio == 0 N = ( 0 if swa_only else (self.max_model_len + self.compress_ratio - 1) // self.compress_ratio ) M = N + self.window_size + self.max_num_batched_tokens shapes = self._prefill_workspace_shapes( M, self.max_num_batched_tokens, q, gather=not self._use_aiter_sparse_mla, ) if shapes: current_workspace_manager().get_simultaneous(*shapes) if mxfp8_out is None: output.zero_() else: mxfp8_out.data.zero_() mxfp8_out.scale.zero_() return assert isinstance(attn_metadata, dict) rocm_metadata = cast( DeepseekV4FlashMLAMetadata | None, attn_metadata.get(self.compressed_cache_prefix) if self.compressed_cache_prefix is not None else None, ) swa_metadata = cast( DeepseekV41ROCMAiterSparseSWAMetadata | None, attn_metadata.get(self.swa_cache_layer.prefix), ) assert swa_metadata is not None swa_only = self.compress_ratio == 0 self_kv_cache = None if swa_only else self._compressed_kv_cache() swa_kv_cache = self.swa_cache_layer.kv_cache num_decodes = swa_metadata.num_decodes num_prefills = swa_metadata.num_prefills num_decode_tokens = swa_metadata.num_decode_tokens if num_prefills > 0: self._forward_prefill( q=q[num_decode_tokens:], positions=positions[num_decode_tokens:], compressed_k_cache=self_kv_cache, swa_k_cache=swa_kv_cache, output=( output[num_decode_tokens:] if isinstance(output, torch.Tensor) else None ), attn_metadata=rocm_metadata, swa_metadata=swa_metadata, mxfp8_out=( None if mxfp8_out is None else ( mxfp8_out.data[num_decode_tokens:], mxfp8_out.scale[num_decode_tokens:], ) ), ) if mxfp8_out is not None: # The reduce epilogue rotates and quantizes every decode row and # the prefill path already did its own, so nothing is owed here. if num_decodes > 0: self._forward_decode( q=q[:num_decode_tokens], positions=positions[:num_decode_tokens], kv_cache=self_kv_cache, swa_metadata=swa_metadata, attn_metadata=rocm_metadata, swa_only=swa_only, output=None, output_mxfp8=( mxfp8_out.data[:num_decode_tokens], mxfp8_out.scale[:num_decode_tokens], ), ) return assert isinstance(output, torch.Tensor) rotated = 0 if num_decodes > 0: rotated = self._forward_decode( q=q[:num_decode_tokens], positions=positions[:num_decode_tokens], kv_cache=self_kv_cache, swa_metadata=swa_metadata, attn_metadata=rocm_metadata, swa_only=swa_only, output=output[:num_decode_tokens], ) # Only the decode reduce rotates its own rows, and only the leading # `rotated` of them; prefill rows and any decode path that did not # fuse still owe the standalone pass. Settle that here rather than in # _o_proj: the split is batch-dependent and _o_proj runs compiled, # where such a value freezes at its trace-time value. rocm_inverse_rope_rows_( output[rotated:, : self.n_local_heads, :], positions[rotated:], self.rotary_emb.cos_sin_cache, self.rope_head_dim, ) def _decode_topk_ragged( self, swa_metadata: DeepseekV41ROCMAiterSparseSWAMetadata, attn_metadata: DeepseekV4FlashMLAMetadata | None, num_decodes: int, num_decode_tokens: int, ) -> _TopkRagged: """Ragged form of the topk indices this layer's index source published. The packing depends only on the shared ``topk_indices_buffer`` and per-step metadata, apart from the layer's own compress ratio, so the source memoizes one result per ratio for all the layers below it. """ assert attn_metadata is not None assert swa_metadata.is_valid_token is not None assert self.topk_indices_buffer is not None assert self._index_source_prefix is not None source = self._static_forward_context[self._index_source_prefix] if source is self: # Fresh indices as of this layer; drop what the last step cached. source._topk_ragged_cache = {} cached = source._topk_ragged_cache.get(self.compress_ratio) if cached is not None: return cached built = compute_global_topk_ragged_indices_and_indptr( self.topk_indices_buffer[:num_decode_tokens], swa_metadata.token_to_req_indices, attn_metadata.block_table[:num_decodes], attn_metadata.block_size // self.compress_ratio, swa_metadata.is_valid_token[:num_decode_tokens], ) source._topk_ragged_cache[self.compress_ratio] = built return built def _forward_decode( self, q: torch.Tensor, positions: torch.Tensor, kv_cache: torch.Tensor | None, swa_metadata: DeepseekV41ROCMAiterSparseSWAMetadata, attn_metadata: DeepseekV4FlashMLAMetadata | None, swa_only: bool, output: torch.Tensor | None, output_mxfp8: tuple[torch.Tensor, torch.Tensor] | None = None, ) -> int: """Returns how many leading rows the decode epilogue inverse-RoPE'd.""" num_decodes = swa_metadata.num_decodes num_decode_tokens = swa_metadata.num_decode_tokens topk_lens = None topk_ragged_indices = None topk_ragged_indptr = None if not swa_only: ( topk_ragged_indices, topk_ragged_indptr, topk_lens, ) = self._decode_topk_ragged( swa_metadata=swa_metadata, attn_metadata=attn_metadata, num_decodes=num_decodes, num_decode_tokens=num_decode_tokens, ) if self._use_aiter_sparse_mla: assert swa_metadata.decode_swa_ragged_indices is not None assert swa_metadata.decode_swa_ragged_indptr is not None out = output if output_mxfp8 is not None: # The aiter kernel has no MXFP8 epilogue: write bf16 rows, then # rotate and quantize them as the decode reduce would have. out = self._mxfp8_bf16_rows(q.shape[0], q) assert out is not None self._aiter_sparse_mla( q, out, self.swa_cache_layer.kv_cache, swa_metadata.decode_swa_ragged_indices, swa_metadata.decode_swa_ragged_indptr, kv_cache, topk_ragged_indices, topk_ragged_indptr, ) if output_mxfp8 is not None: rocm_inverse_rope_mxfp8_rows( out, positions, self.rotary_emb.cos_sin_cache, self.rope_head_dim, output_mxfp8[0], output_mxfp8[1], ) return q.shape[0] # The aiter kernel leaves the inverse RoPE to the caller. return 0 return rocm_sparse_attn_decode( q=q, kv_cache=kv_cache, swa_k_cache=self.swa_cache_layer.kv_cache, swa_only=swa_only, topk_indices=None, topk_lens=topk_lens, swa_indices=swa_metadata.decode_swa_indices, swa_lens=swa_metadata.decode_swa_lens, swa_ragged_indices=swa_metadata.decode_swa_ragged_indices, swa_ragged_indptr=swa_metadata.decode_swa_ragged_indptr, topk_ragged_indices=topk_ragged_indices, topk_ragged_indptr=topk_ragged_indptr, attn_sink=self.attn_sink, scale=self.scale, head_dim=self.head_dim, nope_head_dim=self.nope_head_dim, rope_head_dim=self.rope_head_dim, output=output, output_mxfp8=output_mxfp8, inv_rope_positions=positions, inv_rope_cos_sin_cache=self.rotary_emb.cos_sin_cache, extra_cache_nan_free=_trust_dsv4_extra_cache_nan_free( self.kv_cache_dtype, self._has_kv_transfer, not swa_only and kv_cache is not None, ), ) def _forward_prefill( self, q: torch.Tensor, positions: torch.Tensor, compressed_k_cache: torch.Tensor | None, swa_k_cache: torch.Tensor, output: torch.Tensor | None, attn_metadata: DeepseekV4FlashMLAMetadata | None, swa_metadata: DeepseekV41ROCMAiterSparseSWAMetadata, mxfp8_out: tuple[torch.Tensor, torch.Tensor] | None = None, ) -> None: """Sparse prefill into bf16 ``output``, or into ``mxfp8_out``. With ``mxfp8_out`` the rows go to a bf16 workspace first and are then inverse-RoPE'd and quantized into it in one pass. """ if self._use_aiter_sparse_mla: self._forward_prefill_aiter( q=q, positions=positions, compressed_k_cache=compressed_k_cache, swa_k_cache=swa_k_cache, output=output, attn_metadata=attn_metadata, swa_metadata=swa_metadata, mxfp8_out=mxfp8_out, ) return swa_only = attn_metadata is None num_prefill_tokens = swa_metadata.num_prefill_tokens num_decodes = swa_metadata.num_decodes num_decode_tokens = swa_metadata.num_decode_tokens seq_lens = swa_metadata.prefill_seq_lens gather_lens = swa_metadata.prefill_gather_lens assert seq_lens is not None assert gather_lens is not None # CPU copy, so sizing the gather grid from it does not sync. seq_lens_cpu = swa_metadata.prefill_seq_lens_cpu assert seq_lens_cpu is not None query_start_loc_cpu = swa_metadata.query_start_loc_cpu query_start_loc = swa_metadata.query_start_loc assert query_start_loc_cpu is not None assert query_start_loc is not None prefill_token_base = query_start_loc_cpu[num_decodes] # Local indices filled by the index source; SWA-only layers pass # top_k=0 and never read them. assert self.topk_indices_buffer is not None topk_indices = self.topk_indices_buffer[num_decode_tokens:] topk_indices = topk_indices[:num_prefill_tokens] if not swa_only: top_k = topk_indices.shape[-1] N = (self.max_model_len + self.compress_ratio - 1) // self.compress_ratio else: top_k = 0 N = 0 M = N + self.window_size + self.max_num_batched_tokens chunk_plan = swa_metadata.get_prefill_chunk_plan( compress_ratio=self.compress_ratio, prefill_chunk_size=self.PREFILL_CHUNK_SIZE, has_compressed=not swa_only, ) assert chunk_plan, "prefill chunk plan must be non-empty when num_prefills > 0" workspace = current_workspace_manager().get_simultaneous( *self._prefill_workspace_shapes(M, num_prefill_tokens, q) ) kv_rows = workspace[0].view(-1, q.shape[-1]) if mxfp8_out is not None: output = workspace[1] assert output is not None for chunk_start, chunk_end, chunk_N, chunk_M in chunk_plan: chunk_size = chunk_end - chunk_start kv = kv_rows[: chunk_size * chunk_M].view(chunk_size, chunk_M, -1) if not swa_only: assert attn_metadata is not None assert compressed_k_cache is not None block_table = attn_metadata.block_table[num_decodes:] # compressed_k_cache is OCP on every platform (Triton encoder). dequantize_and_gather_k_cache( kv[:chunk_size], compressed_k_cache, seq_lens=seq_lens[chunk_start:chunk_end] // self.compress_ratio, gather_lens=None, block_table=block_table[chunk_start:chunk_end], block_size=attn_metadata.block_size // self.compress_ratio, offset=0, use_fnuz=False, max_gather_len=int(seq_lens_cpu[chunk_start:chunk_end].max()) // self.compress_ratio, ) swa_block_table = swa_metadata.block_table[num_decodes:] dequantize_and_gather_k_cache( kv[:chunk_size], swa_k_cache, seq_lens=seq_lens[chunk_start:chunk_end], gather_lens=gather_lens[chunk_start:chunk_end], block_table=swa_block_table[chunk_start:chunk_end], block_size=swa_metadata.block_size, offset=chunk_N, use_fnuz=current_platform.is_fp8_fnuz(), ) query_start = ( query_start_loc_cpu[num_decodes + chunk_start] - prefill_token_base ) query_end = ( query_start_loc_cpu[num_decodes + chunk_end] - prefill_token_base ) combined_indices, combined_lens = combine_topk_swa_indices( topk_indices[query_start:query_end], query_start_loc[ num_decodes + chunk_start : num_decodes + chunk_end + 1 ], seq_lens[chunk_start:chunk_end], gather_lens[chunk_start:chunk_end], self.window_size, self.compress_ratio, top_k, chunk_M, chunk_N, ) rocm_sparse_attn_prefill( q=q[query_start:query_end], kv=kv.view(-1, 1, q.shape[-1]), indices=combined_indices, topk_length=combined_lens, scale=self.scale, head_dim=self.head_dim, nope_head_dim=self.nope_head_dim, rope_head_dim=self.rope_head_dim, attn_sink=self.attn_sink, output=output[query_start:query_end], ) if mxfp8_out is not None: rocm_inverse_rope_mxfp8_rows( output, positions[:num_prefill_tokens], self.rotary_emb.cos_sin_cache, self.rope_head_dim, mxfp8_out[0][:num_prefill_tokens], mxfp8_out[1][:num_prefill_tokens], ) def _prefill_workspace_shapes( self, M: int, num_prefill_tokens: int, q: torch.Tensor, gather: bool = True ) -> list[tuple[tuple[int, ...], torch.dtype]]: """The prefill gather buffer, plus bf16 rows for the MXFP8 output. One ``get_simultaneous`` call: separate calls alias the same memory. ``gather=False`` leaves out the gather buffer, which the aiter kernel, reading the caches in place, never needs. """ shapes: list[tuple[tuple[int, ...], torch.dtype]] = [] if gather: shapes.append(((self.PREFILL_CHUNK_SIZE, M, q.shape[-1]), torch.bfloat16)) if _ON_GFX950: shapes.append( ( (num_prefill_tokens, self.n_local_heads, self.head_dim), torch.bfloat16, ) ) return shapes def _prefill_topk_ragged( self, swa_metadata: DeepseekV41ROCMAiterSparseSWAMetadata, attn_metadata: DeepseekV4FlashMLAMetadata, compressed_k_cache: torch.Tensor, ) -> tuple[torch.Tensor, torch.Tensor]: """Prefill counterpart of ``_decode_topk_ragged``: layers that share an index source and a compressed cache reuse one build per step.""" assert swa_metadata.token_to_req_indices is not None assert swa_metadata.query_start_loc is not None assert swa_metadata.seq_lens is not None assert swa_metadata.is_valid_token is not None assert self.topk_indices_buffer is not None assert self._index_source_prefix is not None assert self.compressed_cache_prefix is not None source = self._static_forward_context[self._index_source_prefix] if source is self: source._prefill_topk_ragged_cache = {} cached = source._prefill_topk_ragged_cache.get(self.compressed_cache_prefix) if cached is not None: return cached num_decode_tokens = swa_metadata.num_decode_tokens num_prefill_tokens = swa_metadata.num_prefill_tokens built = build_prefill_topk_ragged_indices( self.topk_indices_buffer[ num_decode_tokens : num_decode_tokens + num_prefill_tokens ], swa_metadata.token_to_req_indices, swa_metadata.query_start_loc, swa_metadata.seq_lens, swa_metadata.is_valid_token, attn_metadata.block_table, block_size=attn_metadata.block_size // self.compress_ratio, compress_ratio=self.compress_ratio, num_compressed=(self.max_model_len + self.compress_ratio - 1) // self.compress_ratio, token_offset=num_decode_tokens, num_rows=compressed_k_cache.shape[0] * compressed_k_cache.shape[1], ) source._prefill_topk_ragged_cache[self.compressed_cache_prefix] = built return built def _forward_prefill_aiter( self, q: torch.Tensor, positions: torch.Tensor, compressed_k_cache: torch.Tensor | None, swa_k_cache: torch.Tensor, output: torch.Tensor | None, attn_metadata: DeepseekV4FlashMLAMetadata | None, swa_metadata: DeepseekV41ROCMAiterSparseSWAMetadata, mxfp8_out: tuple[torch.Tensor, torch.Tensor] | None = None, ) -> None: """Prefill straight off the paged caches, with no bf16 gather. With ``mxfp8_out`` the kernel writes bf16 rows to the workspace, which are then inverse-RoPE'd and quantized into it as ``_forward_prefill`` does. """ num_prefill_tokens = swa_metadata.num_prefill_tokens assert swa_metadata.prefill_swa_ragged_indices is not None assert swa_metadata.prefill_swa_ragged_indptr is not None topk_indices = None topk_indptr = None if attn_metadata is None: compressed_k_cache = None else: assert compressed_k_cache is not None topk_indices, topk_indptr = self._prefill_topk_ragged( swa_metadata, attn_metadata, compressed_k_cache ) if mxfp8_out is not None: output = self._mxfp8_bf16_rows(num_prefill_tokens, q) assert output is not None self._aiter_sparse_mla( q[:num_prefill_tokens], output[:num_prefill_tokens], swa_k_cache, swa_metadata.prefill_swa_ragged_indices, swa_metadata.prefill_swa_ragged_indptr, compressed_k_cache, topk_indices, topk_indptr, ) if mxfp8_out is not None: rocm_inverse_rope_mxfp8_rows( output[:num_prefill_tokens], positions[:num_prefill_tokens], self.rotary_emb.cos_sin_cache, self.rope_head_dim, mxfp8_out[0][:num_prefill_tokens], mxfp8_out[1][:num_prefill_tokens], ) def _mxfp8_bf16_rows(self, num_tokens: int, q: torch.Tensor) -> torch.Tensor: """bf16 attention rows for the aiter kernel to write when the layer's output is MXFP8, taken from the workspace the warmup reserved.""" shapes = self._prefill_workspace_shapes(0, num_tokens, q, gather=False) assert shapes, "an MXFP8 attention output is gfx950-only" return current_workspace_manager().get_simultaneous(*shapes)[0] def _aiter_sparse_mla( self, q: torch.Tensor, output: torch.Tensor, swa_k_cache: torch.Tensor, swa_indices: torch.Tensor, swa_indptr: torch.Tensor, compressed_k_cache: torch.Tensor | None, topk_indices: torch.Tensor | None, topk_indptr: torch.Tensor | None, ) -> None: from vllm._aiter_ops import rocm_aiter_ops # Both fp8 caches are read in place: the SWA window as the main segment, # the top-k compressed tokens as the extra one. rocm_aiter_ops.triton_sparse_mla_fwd( q, swa_k_cache, output, self.scale, swa_indptr, swa_indices, attn_sink=self.attn_sink[: q.shape[1]], extra_kv_buffer=compressed_k_cache, extra_kv_indptr=topk_indptr, extra_kv_indices=topk_indices, )